目录
Page磁盘读取过程
buf_page_get_gen
| | ==> rw_lock_s_lock(hash_lock);
| | ==> block = (buf_block_t*) buf_page_hash_get_low(buf_pool, space, offset, fold);
| | ==> if (block == NULL) rw_lock_s_unlock(hash_lock);
| | ==> //(MM)从存储读取PAGE到BP
| | ==> buf_read_page(space, zip_size, offset)
| | ==> count = buf_read_page_low(&err, true, BUF_READ_ANY_PAGE, space,
zip_size, FALSE, tablespace_version, offset);
| | ==> //从BP中申请空闲page空间
| | ==> bpage = buf_page_init_for_read(err, mode, space, zip_size, unzip, tablespace_version, offset);
| | ==> block = NULL;
| | ==> buf_pool_mutex_enter(buf_pool);
| | ==> rw_lock_x_lock(hash_lock);
| | ==> watch_page = buf_page_hash_get_low(buf_pool, space, offset, fold);
| | ==> //释放了hashlock,还持有BP lock
| | ==> rw_lock_x_unlock(hash_lock);
| | ==> //申请空闲内存块
| | ==> data = buf_buddy_alloc(buf_pool, zip_size, &lru);
| | ==> buf_buddy_alloc_low(buf_pool, buf_buddy_get_slot(size), lru));
| | ==> //从Free List中获取内存块
| | ==> block = buf_LRU_get_free_only(buf_pool);
| | ==> //free list中没有,从LRU中获取,此时要释放BP锁,因为从LRU获取过程中可能会休眠
| | ==> buf_pool_mutex_exit(buf_pool);
| |
| |
| | ==> block = buf_LRU_get_free_block(buf_pool);
| | ==> //标识从LRU链表中获取
| | ==> *lru = TRUE;
| | ==> //获取成功后重新加锁
| | ==> buf_pool_mutex_enter(buf_pool);
| | ==> rw_lock_x_lock(hash_lock);
| | ==> //如果从LRU获取的话,需要重新CHECK hash表中是否有该Page,因为获取过程中会释放BP MUTEX
| | ==> /* If buf_buddy_alloc() allocated storage from the LRU list, it released and reacquired buf_pool->mutex. Thus, we must check the page_hash again, as it may have been modified. */
| | ==> if (UNIV_UNLIKELY(lru)) {
| | ==> watch_page = buf_page_hash_get_low(buf_pool, space, offset, fold);
| | ==> if (watch_page && !buf_pool_watch_is_sentinel(buf_pool, watch_page)) rw_lock_x_unlock(hash_lock); watch_page = NULL; buf_buddy_free(buf_pool, data, zip_size);
| | ==> //(MM)插入到HASH表
| | ==> HASH_INSERT(buf_page_t, hash, buf_pool->page_hash, fold, bpage);
| | ==> rw_lock_x_unlock(hash_lock);
| | ==> //(MM)插入到LRU链表
| | ==> buf_LRU_add_block(bpage, TRUE/* to old blocks */);
| | ==> buf_pool_mutex_exit(buf_pool);
| | ==> //增加引用计数
| | ==> buf_block_fix(fix_block);
| | ==> #ifdef PAGE_ATOMIC_REF_COUNT
| | ==> os_atomic_increment_uint32(&block->page.buf_fix_count, 1);
| | ==> #else
| | ==> ib_mutex_t* block_mutex = buf_page_get_mutex(&block->page);
| | ==> mutex_enter(block_mutex);
| | ==> ++block->page.buf_fix_count;
| | ==> mutex_exit(block_mutex);
| | ==> rw_lock_s_unlock(hash_lock);
| | ==> buf_wait_for_read(fix_block);
过程是:
- 对Page_Hash加s锁,判断是否存在。
- 如果不存在,释放s锁,加bp的互斥锁、加x锁。再次判断是否存在。
- 如果不存在,释放x锁,申请空闲Block。
- 然后加x锁,判断是否存在,如果不存在,插入空闲BLock。释放x锁
- 把block加入lru链表。释放bp的互斥锁。
步骤3释放x锁的原因是,线程持有bp的互斥锁,对lru\freelist的操作都依靠互斥锁。因此可以通过bp的互斥锁,阻塞其他线程申请空闲Block。但是在从Lru申请BLock过程中,可能出现由于休眠等待需要释放互斥锁,而导致其他线程进入,因此第4步根据空闲BLock的来源判断是否需要重新Check。
可以看到,从磁盘读取过程中,对hash的互斥成本是比较高的,加了2次x锁。这个地方后续版本中有优化。
申请空闲PAGE空间
buf_LRU_get_free_block
| | ==> buf_pool_mutex_enter(buf_pool);
| | ==> //(MM)从free_list中获取Page空间
| | ==> block = buf_LRU_get_free_only(buf_pool);
| | ==> UT_LIST_REMOVE(list, buf_pool->free, (&block->page));
| | ==> //(MM)如果没有获取到,从LRU链表中获取非脏的PAGE空间
| | ==> freed = buf_LRU_scan_and_free_block(buf_pool, n_iterations > 0);
| | ==> //(MM)如果没有,从LLU刷一个脏页下去
| | ==> buf_flush_single_page_from_LRU(buf_pool)
buf_LRU_scan_and_free_block(buf_pool, n_iterations > 0)
| | ==> //尝试从解压缩页中获取空间
| | ==> buf_LRU_free_from_unzip_LRU_list(buf_pool, scan_all)
| | ==> //从LRU中获取非脏页面
| | ==> buf_LRU_free_from_common_LRU_list(buf_pool, scan_all)
| | ==> //从后向前遍历LRU链表
| | ==> for (bpage = UT_LIST_GET_LAST(buf_pool->LRU), scanned = 1, freed = FALSE;
bpage != NULL && !freed && (scan_all || scanned < srv_LRU_scan_depth); ++scanned) {
| | ==> accessed = buf_page_is_accessed(bpage);
| | ==> //尝试从LRU链表中转移一个PAGE到FREE LIST
| | ==> freed = buf_LRU_free_page(bpage, true);
| | ==> //判断页面是否有引用或者在IO过程中
| | ==> if (!buf_page_can_relocate(bpage)) exit
| | ==> //判断页面是否有修改
| | ==> if (bpage->oldest_modification) exit
| | ==> //从LRU链表中删除一条PAGE空间,如果是BUF_BLOCK_FILE_PAGE类型的PAGE,要在外面调用函数加入FREELIST
| | ==> buf_LRU_block_remove_hashed(bpage, zip)
| | ==> buf_LRU_remove_block(bpage);
| | ==> UT_LIST_REMOVE(LRU, buf_pool->LRU, bpage);
| | ==> //从HASH表中删除
| | ==> HASH_DELETE(buf_page_t, hash, buf_pool->page_hash, fold, bpage);
| | ==> //把Page加入到FREE LIST
| | ==> buf_LRU_block_free_hashed_page((buf_block_t*) bpage);
| | ==> buf_LRU_block_free_non_file_page(block);
| | ==> memset
| | ==> UT_LIST_ADD_FIRST(list, buf_pool->free, (&block->page));
| | ==> }
buf_flush_single_page_from_LRU(buf_pool)
| | ==> for (bpage = UT_LIST_GET_LAST(buf_pool->LRU), scanned = 1;
bpage != NULL;
bpage = UT_LIST_GET_PREV(LRU, bpage), ++scanned) {
| | ==> //判断页面是否有修改且是否在IO中
| | ==> if (buf_flush_ready_for_flush(bpage, BUF_FLUSH_SINGLE_PAGE)
| | ==> //刷脏页(同步IO)
| | ==> buf_flush_page(buf_pool, bpage, BUF_FLUSH_SINGLE_PAGE, true);
| | ==> buf_page_set_io_fix(bpage, BUF_IO_WRITE);
| | ==> //如果是BUF_FLUSH_LIST,可能会走下面函数,此处有三个且条件
| | ==> buf_dblwr_flush_buffered_writes();
| | ==> buf_flush_write_block_low(bpage, flush_type, sync);
| | ==> if (flush_type == BUF_FLUSH_SINGLE_PAGE) buf_dblwr_write_single_page(bpage, sync);
| | ==> //持久化dblwr
| | ==> fil_flush(TRX_SYS_SPACE);
| | ==> //持久化数据页
| | ==> buf_dblwr_write_block_to_datafile(bpage, sync);
| | ==> //调IO接口
| | ==> fil_io(flags, sync, buf_block_get_space(block), 0,
buf_block_get_page_no(block), 0, UNIV_PAGE_SIZE,
(void*) block->frame, (void*) block)
| | ==> //调AIO接口
| | ==> ret = os_aio(type, mode | wake_later, node->name, node->handle, buf, offset, len, node, message);
| | ==> os_aio_func
| | ==> //如果同步IO,读调用os_file_read_func,写调用os_file_write_func
| | ==> //如果异步IO,走异步IO系统
| | ==> os_aio_simulated_wake_handler_thread
| | ==> //如果是同步IO,进行IO_WAIT
| | ==> fil_node_complete_io(node, fil_system, type);
| | ==> //关闭DBLWR场景
| | ==> fil_io
| | ==> //其他场景,通过dblwr持久化数据页
| | ==> buf_dblwr_add_to_batch(bpage);
| | ==> //同步IO场景,执行IO完成后操作
| | ==> if (sync) {
| | ==> fil_flush(buf_page_get_space(bpage));
| | ==> buf_page_io_complete(bpage);
| | ==> buf_flush_write_complete(bpage)
| | ==> //从FLU链表中移出
| | ==> buf_flush_remove(bpage);
| | ==> }
| | ==> }
| | ==> //因为刚刷完脏页,重新遍历LRU链表,尝试把转移一个PAGE到FREE_LIST
| | ==> ready = buf_flush_ready_for_replace(bpage);
| | ==> freed = buf_LRU_free_page(bpage, evict_zip);
刷脏页和LRU链表
buf_flush_page_cleaner_thread
| | ==> //依次遍历每个Buffer pool instance,从LRU尾部开始扫描,直到第srv_LRU_scan_depth个page停止,
| | ==> //按批次刷LRU,每次期望刷100个page(PAGE_CLEANER_LRU_BATCH_CHUNK_SIZE), 每个Bp会进行srv_LRU_scan_depth/100次循环
| | ==> buf_flush_LRU_tail
| | ==> for (ulint i = 0; i < srv_buf_pool_instances; i++) {
| | ==> scan_depth = UT_LIST_GET_LEN(buf_pool->LRU);
| | ==> //LRU释放PAGE到FREE LIST
| | ==> buf_flush_LRU(buf_pool, PAGE_CLEANER_LRU_BATCH_CHUNK_SIZE, &n_flushed)
| | ==> buf_flush_start(buf_pool, BUF_FLUSH_LRU)
| |
| |
| | ==> buf_flush_batch(buf_pool, BUF_FLUSH_LRU, min_n, 0);
| | ==> case BUF_FLUSH_LRU: buf_do_LRU_batch(buf_pool, min_n);
| | ==> if (buf_LRU_evict_from_unzip_LRU(buf_pool)) buf_free_from_unzip_LRU_list_batch(buf_pool, max);
| | ==> //
| | ==> buf_flush_LRU_list_batch(buf_pool, max - count);
| | ==> //没有引用、没有IO、没有变更
| | ==> evict = buf_flush_ready_for_replace(bpage);
| | ==> //尝试从LRU链表中转移PAGE到FREE LIST
| | ==> if (evict) buf_LRU_free_page(bpage, true);
| | ==> //把这个PAGE持久化
| | ==> else buf_flush_page_and_try_neighbors(bpage, BUF_FLUSH_LRU, max, &count);
| | ==> //判断这个PAGE是否需要持久化
| | ==> buf_flush_ready_for_flush(bpage, flush_type)
| | ==> //持久化这个PAGE
| | ==> buf_flush_try_neighbors(space, offset, flush_type, *count, n_to_flush)
| | ==> buf_flush_page(buf_pool, bpage, flush_type, false)
| | ==> buf_flush_end(buf_pool, BUF_FLUSH_LRU);
| | ==> }
| | ==> //判断是需要刷脏页
| | ==> page_cleaner_flush_pages_if_needed
| | ==> //number of pages thatwe should attempt to flush, LSN up to which flushing must happen
| | ==> page_cleaner_do_flush_batch(ulint n_to_flush, lsn_t lsn_limit)
| | ==> buf_flush_list(n_to_flush, lsn_limit, &n_flushed);
| | ==> //依次为每个BP执行
| | ==> buf_flush_batch(buf_pool, BUF_FLUSH_LIST, min_n, lsn_limit)
| | ==> count = buf_do_flush_list_batch(buf_pool, min_n, lsn_limit);
| | ==> //从FLU链表的尾部开始遍历,跳出条件时候刷脏的PAGE数量和LSN
| | ==> buf_flush_set_hp(buf_pool, prev);
| | ==> buf_flush_list_mutex_exit(buf_pool);
| | ==> buf_flush_page_and_try_neighbors(bpage, BUF_FLUSH_LIST, min_n, &count);
| | ==> buf_flush_list_mutex_enter(buf_pool);
| | ==> buf_flush_is_hp(buf_pool, prev)
Page内存读取和Page淘汰的互斥
buf_flush_batch
| | ==> //加LRU LOCK
| | ==> buf_pool_mutex_enter(buf_pool);
| | ==> count = buf_do_LRU_batch(buf_pool, min_n);
| | ==> buf_flush_LRU_list_batch(buf_pool, max - count);
| | ==> bpage = UT_LIST_GET_LAST(buf_pool->LRU);
| | ==> mutex_enter(block_mutex);
| | ==> evict = buf_flush_ready_for_replace(bpage);
| | ==> bpage->buf_fix_count == 0
| | ==> mutex_exit(block_mutex);
| | ==> if (evict) buf_LRU_free_page
| | ==> rw_lock_x_lock(hash_lock);
| | ==> mutex_enter(block_mutex);
| | ==> //重新判断引用计数
| | ==> buf_page_can_relocate
| | ==> //从LRU和HASH表删除
| | ==> buf_LRU_block_remove_hashed
| | ==> buf_LRU_remove_block(bpage);
| | ==> HASH_DELETE(buf_page_t, hash, buf_pool->page_hash, fold, bpage);
| | ==> rw_lock_x_unlock(hash_lock);
| | ==> buf_pool_mutex_exit(buf_pool);
| | ==> buf_pool_mutex_enter(buf_pool);
| | ==> buf_LRU_block_free_hashed_page((buf_block_t*) bpage);
| | ==> buf_LRU_block_free_non_file_page(block);
| | ==> UT_LIST_ADD_FIRST(list, buf_pool->free, (&block->page));
| | ==> buf_pool_mutex_exit(buf_pool);
Page淘汰过程如下:
- 加BP的互斥锁
- 从LRU链表的获取最老的数据页
- 确认Page是否可以被淘汰
- 如果可以淘汰,加hash_page的x锁,此时重新判断,因为前述判断无法拦截对Page的访问,因此此时对Page_hash加x锁,进行准确的判断。
- 如果确实可以淘汰,从Page_Hash中删除,释放hash的x锁。
- 将Page加入空闲链表